ZFCRADO.PDF

A Statement in Combinatorics that is
Independent of ZFC
by Stephen Fenner1 and William Gasarch2
1
Introduction
There are some statements that are independent of Zermelo-Frankl Set Theory (henceforth
ZFC). Such statements cannot be proven or disproven by conventional mathematics. The
Continuum Hypothesis is one such statement (“There is no cardinality strictly between ℵ0
and 2ℵ0 .”) Many such statements are unnatural in that they deal with objects only set
theorists and other logicians care about.
We present a natural statement in combinatorics that is independent of ZFC. The result
is due to Erdős. In the last section we will discuss the question of whether the statement is
really natural.
Notation 1.1. We use N to denote {0, 1, 2, . . .}. We use N+ to denote {1, 2, 3, . . .}. If
n ∈ N+ then [n] is the set {1, 2, . . . , n}. We use R and C to denote the sets of real and
complex numbers, respectively. We use Z to denote the integers. We use k-AP to refer
to an arithmetic progression with k distinct elements. For a set A and k ∈ N, we let A
k
denote the set of k-element subsets of A.
2
Colorings and equations
Definition 2.1. A finite coloring of a set S is a map from S to a finite set. An ℵ0 -coloring
of a set S is a map from S to a countable set.
The following theorem is well known. We prove it for the sake of completeness.
Theorem 2.2. For any finite coloring of N+ , there exists distinct monochromatic e1 , e2 , e3 , e4
such that
e1 + e2 = e3 + e4 .
Proof. Let COL be a finite coloring of N+ . Let [c] be image of COL.
First Proof
Recall Ramsey’s theorem (6; 8; 13) on N: for any finite coloring of unordered pairs of
naturals there exists an infinite set A such that all pairs of elements from A have the same
color.
Let COL∗ : N2 → [c] be defined by COL∗ ({a, b}) = COL(|a − b|). Let A be the set that
exists by Ramsey’s theorem. Let a1 < n
a2 < a3 < a4 ∈ A. Since
A is infinite we can take
o
[4]
a1 , a2 , a3 , a4 such that the six numbers aj − ai | {i, j} ∈ 2
are distinct.
1
2
[email protected]
[email protected]
1
Since all of the COL∗ ({ai , aj }) are the same color we have that, for i < j, COL(aj − ai )
are all the same color. Let
e1 = a2 − a1
e2 = a4 − a2
e3 = a3 − a1
e4 = a4 − a3
Clearly e1 , e2 , e3 , e4 are distinct, COL(e1 ) = COL(e2 ) = COL(e3 ) = COL(e4 ), and
e1 + e2 = e3 + e4 .
Second Proof
Recall van der Waerden’s theorem (7; 8; 10; 15): For all k, for any finite coloring of
N+ , there exists a monochromatic k-AP, that is, a k-AP all of whose elements are the same
color.
Apply van der Waerden’s Theorem to COL with k = 4. There exists a, d ∈ N+ such
that a, a + d, a + 2d, a + 3d are the same color. Let
e1 = a
e2 = a + 4d
e3 = a + 2d
e4 = a + 3d
Note 2.3. Rado’s theorem characterizes which equations lead to theorems like Theorem 2.2 and which ones do not. We will discuss Rado’s theorem in Section 7.
3
What if we color the reals?
What if we finitely color the Reals? Theorem 2.2 will still hold since we can just restrict
the coloring to N+ . What if we ℵ0 -color the reals?
Let S be the following statement:
For any ℵ0 -coloring of the reals, there exist distinct monochromatic e1 , e2 , e3 , e4 such
that
e1 + e2 = e3 + e4 .
Is S true? Before answering this we need to decide what truth really is. We consider a
statement to be true iff it can be derived from the axioms of ZFC (Zermelo-Frankel with
Choice). This encompasses most of mathematics. All of our reasoning is done in ZFC.
It turns out that S is equivalent to the negation of CH, and hence is independent of ZFC.
Komjáth (9) claims that Erdős proved this result. The proof we give is due to Davies (3).
The goal of our paper is to present and popularize this result.
2
Definition 3.1. ω is the first infinite ordinal, namely {1 < 2 < 3 < · · · }. (Formally it is
any ordering that is equivalent to {1 < 2 < 3 < · · · }.) ω1 is the first uncountable ordinal.
ω2 is the first ordinal with cardinality bigger than ω1 .
Fact 3.2.
1. If CH is true, then there is a bijection between R and ω1 . For all α ∈ ω1 let α map to
xα . We can picture the reals listed out as such:
x0 , x1 , x2 , . . . , xα , . . . .
Note that, for all α ∈ ω1 , the set {xβ | β < α} is countable.
2. If CH is false, then there is an injection from ω2 to R.
4
CH =⇒ ¬S
Definition 4.1. Let X ⊆ R. Then CL(X) is the smallest set Y ⊇ X of reals such that
a, b, c ∈ Y
=⇒ a + b − c ∈ Y.
Lemma 4.2.
1. If X is countable then CL(X) is countable.
2. If X1 ⊆ X2 then CL(X1 ) ⊆ CL(X2 ).
Proof. 1) Assume X is countable. CL(X) can be defined with an ω-induction (that is, an
induction just through ω).
C0 = X
Cn+1 = Cn ∪ {a + b − c | a, b, c ∈ Cn }
One can easily show that CL(X) = ∪∞
i=0 Ci and that this set is countable.
2) This is an easy exercise.
Theorem 4.3. Assume CH is true. There exists an ℵ0 -coloring of R such that there are
no distinct monochromatic e1 , e2 , e3 , e4 such that
e1 + e2 = e3 + e4 .
Proof. Since we are assuming CH is true, we have, by Fact 3.2.1, a bijection between R and
ω1 . For each α ∈ ω1 let xα be the real that α maps to.
For α < ω1 let
Xα = {xβ | β < α}.
Note the following:
3
1. For all α, Xα is countable.
2. X0 ⊂ X1 ⊂ X2 ⊂ X3 ⊂ · · · ⊂ Xα ⊂ · · ·
S
3. α<ω1 Xα = R.
We define another increasing sequence of sets Yα by letting
Yα = CL(Xα ).
Note the following:
1. For all α, Yα is countable. This is from Lemma 4.2.1.
2. Y0 ⊆ Y1 ⊆ Y2 ⊆ Y3 ⊆ · · · ⊆ Yα ⊆ · · · . This is from Lemma 4.2.2.
S
3. α<ω1 Yα = R.
We now define our last sequence of sets:
For all α < ω1 ,

Zα = Yα − 

[
Yβ  .
β<α
Note the following:
1. Each Zα is finite or countable.
2. The Zα form a partition of R (although some of the Zα may be empty).
We will now define an ℵ0 -coloring of R: For each α < ω1 we color Zα with colors in ω
making sure that every element of Zα has a different color (this is possible since Zα is at
most countable).
Assume, by way of contradiction, that there are distinct monochromatic e1 , e2 , e3 , e4
such that
e1 + e2 = e3 + e4 .
Let α1 , α2 , α3 , α4 ∈ ω1 be such that ei ∈ Zαi . Since all of the elements in any Zα are
colored differently, all of the αi ’s are different. We will assume α1 < α2 < α3 < α4 . The
other cases are similar. Note that
e4 = e1 + e2 − e3 .
and
e1 , e2 , e3 ∈ Zα1 ∪ Zα2 ∪ Zα3 ⊆ Yα1 ∪ Yα2 ∪ Yα3 = Yα3 .
Since Yα3 = CL(Xα3 ) and e1 , e2 , e3 ∈ Yα3 , we have e4 ∈ Yα3 . Hence e4 ∈
/ Zα4 . This is a
contradiction.
What was it about the equation
e1 + e2 = e3 + e4
that made the proof of Theorem 4.3 work? Absolutely nothing:
4
Theorem 4.4. Let n ≥ 2. Let a1 , . . . , an ∈ R be nonzero. Assume CH is true. There exists
an ℵ0 -coloring of R such that there are no distinct monochromatic e1 , . . . , en such that
n
X
ai ei = 0.
i=1
Proof sketch. Since this proof is similar to the last one we just sketch it.
Definition 4.5. Let X ⊆ R. CL(X) is the smallest superset of X such that the following
holds:
For all m ∈ {1, . . . , n} and for all e1 , . . . , em−1 , em+1 , . . . , en ,
X
e1 , . . . , em−1 , em+1 , . . . , en ∈ CL(X) =⇒ −(1/am )
ai ei ∈ CL(X).
i∈{1,...,n}−{m}
Let Xα , Yα , Zα be defined as in Theorem 4.3 using this new definition of CL. With
these definitions define an ℵ0 -coloring like the one in the proof of Theorem 4.3.
Assume, by way of contradiction, that there are distinct monochromatic e1 , . . . , en such
that
n
X
ai ei = 0.
i=1
Let α1 , . . . , αn be such that ei ∈ Zαi . Since all of the elements in any Zα are colored
differently, all of the αi ’s are different. We will assume α1 < α2 < · · · < αn . The other
cases are similar. Note that
en = −(1/an )
n−1
X
ai ei ∈ CL(X)
i=1
and
e1 , . . . , en−1 ∈ Zα1 ∪ · · · ∪ Zαn−1 ⊆ Yαn−1 .
Since Yαn−1 = CL(Xαn−1 ) and e1 , . . . , en−1 ∈ Yαn−1 , we have en ∈ Yαn−1 . Hence en ∈
/
Zαn . This is a contradiction.
Note 4.6. The converse to Theorem 4.4 is not true. The s = 2 case of Theorem 7.7
(in Section 7) states that every ℵ0 -coloring of N has a distinct monochromatic solution
to x1 + 2x2 = x3 + x4 + x5 is ℵ0 iff 2ℵ0 > ℵ2 . Therefore if 2ℵ0 = ℵ2 (so CH is false)
then there is an ℵ0 -coloring of N such that there is no monochromatic distinct solution to
x1 + 2x2 = x3 + x4 + x5 . This contradicts the converse of Theorem 4.4
5
¬ CH =⇒ S
Theorem 5.1. Assume CH is false. For any ℵ0 -coloring of R there exist distinct monochromatic e1 , e2 , e3 , e4 such that
e1 + e2 = e3 + e4 .
5
Proof. By Fact 3.2 there is an injection of ω2 into R. If α ∈ ω2 , then xα is the real associated
to it.
Given an ℵ0 -coloring COL of R we show that there exist distinct monochromatic e1 , e2 , e3 , e4
such that e1 + e2 = e3 + e4 .
We define a map F from ω2 to ω1 × ω1 × ω1 × ω as follows:
1. Let β ∈ ω2 .
2. Define a map from ω1 to ω by
α 7→ COL(xα + xβ ).
3. Let α1 , α2 , α3 ∈ ω1 be distinct elements of ω1 , and i ∈ ω, such that α1 , α2 , α3 all map
to i. Such α1 , α2 , α3 , i clearly exist since ℵ0 + ℵ0 = ℵ0 < ℵ1 . (There are ℵ1 many
elements that map to the same element of ω, but we do not need that.)
4. Map β to (α1 , α2 , α3 , i).
Since F maps a set of cardinality ℵ2 to a set of cardinality ℵ1 , there exists some element
that is mapped to twice by F (actually there is an element that is mapped to ℵ2 times, but
we do not need this). Let α1 , α2 , α3 be distinct elements of ω1 , i ∈ ω, and β, β 0 be distinct
elements of ω2 , such that
F (β) = F (β 0 ) = (α1 , α2 , α3 , i).
Choose distinct α, α0 ∈ {α1 , α2 , α3 } such that xα − xα0 ∈
/ {xβ − xβ 0 , xβ 0 − xβ }. We can
do this because there are at least three possible values for xα − xα0 .
Since F (β) = (α1 , α2 , α3 , i), we have
COL(xα + xβ ) = COL(xα0 + xβ ) = i.
Since
F (β 0 )
= (α1 , α2 , α3 , i), we have
COL(xα + xβ 0 ) = COL(xα0 + xβ 0 ) = i.
Let
e1 = xα + xβ
e2 = xα0 + xβ 0
e3 = xα0 + xβ
e4 = xα + xβ 0 .
Then
COL(e1 ) = COL(e2 ) = COL(e3 ) = COL(e4 ) = i
and
e1 + e2 = e3 + e4 = xα + xα0 + xβ + xβ 0 .
Since xα 6= xα0 and xβ 6= xβ 0 , we have {e1 , e2 } ∩ {e3 , e4 } = ∅.
Moreover, the equation e1 = e2 is equivalent to
xα − xα0 = xβ 0 − xβ ,
which is ruled out by our choice of α, α0 , and so e1 6= e2 .
Similarly, e3 6= e4 .
Thus e1 , e2 , e3 , e4 are all distinct.
6
6
A generalization
All the results of the last two sections about ℵ0 -colorings of R hold practically verbatim with
R replaced by Rk , for any fixed integer k ≥ 1. In this more geometrical context, e1 , e2 , e3 , e4
are vectors in k-dimensional Euclidean space, and the equation e1 + e2 = e3 + e4 says that
e1 , e2 , e3 , e4 are the vertices of a parallelogram (whose area may be zero). In particular, we
have the following two theorems:
Theorem 6.1. Fix any integer k ≥ 1. The following are equivalent:
1. 2ℵ0 > ℵ1 .
2. For any ℵ0 -coloring of Rk , there exist distinct monochromatic vectors e1 , e2 , e3 , e4 ∈ Rk
such that e1 + e2 = e3 + e4 .
Theorem 6.2. Fix any integers k ≥ 1 and n ≥ 2, and let a1 , . . . , , an ∈ R be nonzero. If CH
is true, then there exists an ℵ0 -coloring of Rk such that there are no distinct monochromatic
vectors e1 , . . . , en ∈ Rk such that
n
X
ai ei = 0.
i=1
7
More is known
Theorem 2.2 is a special case of a general theorem about colorings and equations.
Definition 7.1.
Let ~b = (b1 , . . . , bn ) ∈ Zn
1. ~b is regular if the following holds: For all finite colorings of N+ there exist monochromatic e1 , . . . , en ∈ N+ such that
n
X
bi ei = 0.
i=1
2. ~b is distinct regular if the following holds: For all finite colorings of N+ there exist
monochromatic e1 , . . . , en ∈ N+ , all distinct, such that
n
X
bi ei = 0.
i=1
In 1916 Schur (14) (see also (7; 8)) proved that, for any finite coloring of N+ , there is a
monochromatic solution to x+y = z. Using the above terminology he proved that (1, 1, −1)
was regular. For him this was a Lemma en route to an alternative proof to the following
theorem of Dickson (4):
For all n ≥ 2 there is a prime p0 such that, for all primes p ≥ p0 , xn + y n = z n has a
nontrivial solution mod p.
For an English version of Schur’s proof of Dickson’s theorem see either the book by
Graham-Rothschild-Spencer (8) or the free online book by Gasarch-Kruskal-Parrish (7).
Schur’s student Rado (11; 12) (see also (8; 7)) proved the following generalization of
Schur’s lemma:
7
Theorem 7.2.
1. ~b is regular iff some subset of b1 , . . . , bn sums to 0.
2. ~b is distinct-regular iff some subset of b1 , . . . , bn sums to 0 and there exists a vector ~λ
of distinct reals such that ~b · ~λ = 0.
Note 7.3. Rado’s theorem is about any finite coloring. What about any (say) 3-coloring?
An equation is k-regular if for any k-coloring of N there is a monochromatic solution. There
is no known characterization of which equations are k-regular. Alexeev and Tsimmerman (1)
have shown that there are equations that are (k − 1)-regular that are not k-regular.
We want to summarize the equivalence of S and ¬CH using the notion of regularity.
Definition 7.4. ~b is ℵ0 -distinct regular if the following holds: For all ℵ0 -colorings of R
there exist distinct monochromatic e1 , . . . , en ∈ R such that
n
X
bi ei = 0.
(1)
i=1
Notation 7.5. We may also say that an equation is ℵ0 -distinct-regular. For example, the
statement x1 + x2 = x3 + x4 is ℵ0 -distinct-regular means that (1, 1, −1, −1) is ℵ0 -distinctregular.
If we combine Theorems 4.3 and 5.1 and use this definition of regular we obtain the
following.
Theorem 7.6. x1 + x2 = x3 + x4 is ℵ0 -distinct-regular iff 2ℵ0 > ℵ1 .
What about other linear equations over the reals? Jacob Fox (5) has generalized Theorem 7.6 to prove the following.
Theorem 7.7. Let s ∈ N. The equation
x1 + sx2 = x3 + · · · + xs+3
(2)
is ℵ0 -distinct regular iff 2ℵ0 > ℵs .
Since we are considering coloring the reals, one can easily define what it means for
tuples (b1 , . . . , bn ) ∈ Rn or other domains to be ℵ0 -distinct regular. A theorem of Ceder
(2, Theorem 4) gives us some information about when (b1 , b2 , b3 ) is ℵ0 -distinct regular for
arbitrary complex b1 , b2 , b3 (not just integers). It uses no assumptions outside of ZFC.
Theorem 7.8 (Ceder). If P is any polygon with at least 3 vertices in R1 (or R2 ), then
there exists an ℵ0 -coloring of R1 (resp R2 ) and such that there is no monochromatic polygon
similar to P .
Corollary 7.9. For any b1 , b2 , b3 ∈ C not all zero, if b1 + b2 + b3 = 0, then (b1 , b2 , b3 ) is not
ℵ0 -distinct regular.
8
Proof. Suppose b1 + b2 + b3 = 0 and at least one of the bi is nonzero. We can assume all
the bi are nonzero. (If any of the bi is zero—say b1 —then Equation (1) reduces to e2 = e3 ,
which cannot be satisfied with distinct e2 , e3 .) Dividing by b3 and rearranging, Equation (1)
becomes
e3 − e1 = γ(e2 − e1 ),
(3)
where γ := −b2 /b3 . Note that γ ∈
/ {0, 1}, which forces the ei to be distinct. Identify C with
R2 in the usual way. Any e1 , e2 , e3 ∈ C satisfying (3) form the vertices of a triangle similar
to the triangle P = {0, 1, γ}. By Theorem 7.8, we get an ℵ0 -coloring of C—based on the
partition {An }∞
n=1 —that monocolors no such {e1 , e2 , e3 }.
There is nothing special about the field C of Ceder’s result. Any field will do.
Theorem 7.10. Let F be any field. For any γ ∈ F − {0, 1}, there exists an ℵ0 -coloring of
F such that there are no distinct monochromatic x, y, z ∈ F such that
z − x = γ(y − x).
(4)
Proof. This proof is essentially Ceder’s, generalized to F . Let K be some countable subfield
of F containing γ. Choose a basis {bi }i∈I of F as a vector space over K, where I is some
index set with a linear order <. Then for any w ∈ F , there are unique coordinates {wi }i∈I
where each wi ∈ K, only finitely many of the wi are nonzero, and
X
w=
wi bi .
i∈I
Define the support of w as
supp(w) := {i ∈ I : wi 6= 0} = {i1 < i2 < · · · < ik }
for some k. Then define the signature of w as the k-tuple of the nonzero coordinates of w,
namely
sig(w) := (wi1 , wi2 , . . . , wik ).
Note that supp(w) and sig(w) together uniquely determine w. Also note that there are
only countably many possible signatures, so coloring each element of F by its signature
gives us an ℵ0 -coloring of F .
Suppose x, y, z ∈ F are distinct and satisfy Equation (4). We show that x, y, and z
cannot all have the same signature. Equation (4) is equivalent to
z = γy + (1 − γ)x,
or equivalently, since γ ∈ K,
(∀i ∈ I)[ zi = γyi + (1 − γ)xi ].
If sig(x) 6= sig(y), then we are done, so assume sig(x) = sig(y). Since x 6= y, we must have
supp(x) 6= supp(y). Let ` ∈ I be the least element of supp(x) 4 supp(y). Then for every
j < `, we have xj = yj , and so
zj = γyj + (1 − γ)xj = yj = xj .
We now have two cases for `:
9
Case 1: ` ∈ supp(y). Then y` 6= 0 and x` = 0, because ` ∈
/ supp(x). This gives
z` = γy` ∈
/ {0, y` },
which puts ` into supp(z) and forces sig(z) 6= sig(y).
Case 2: ` ∈ supp(x). A similar argument, swapping the roles of x and y and swapping γ
with 1 − γ, shows that sig(z) 6= sig(x).
Corollary 7.11. Let F be any field. For any b1 , b2 , b3 ∈ F not all zero, if b1 + b2 + b3 = 0,
then (b1 , b2 , b3 ) is not ℵ0 -distinct regular.
8
Is the statement really natural?
Theorem 4.3 and 5.1 are stated as though they are about R. However, all that is used about
R is that it is a vector space over Q. Hence the proof we gave really proves Theorem 8.2
below, from which Theorems 4.3 and 5.1 (as well as Theorems 6.1 and 6.2, for that matter)
follow as easy corollaries.
Definition 8.1. For any vector space V over Q, let S(V ) be the statement, “For any ℵ0 coloring of V there exist distinct monochromatic e1 , e2 , e3 , e4 ∈ V such that e1 +e2 = e3 +e4 .”
Theorem 8.2. If V is a vector space over Q, then S(V ) iff V has dimension at least ℵ2 .
The proof of Theorem 8.2 is in ZFC.
One can ask the following: Since the result, when abstracted, has nothing to do with
R and is just a statement provable in ZFC, do we really have a natural statement that is
independent of ZFC? We believe so.
After you know that every finite coloring of N has a distinct monochromatic solution to
e1 + e2 = e3 + e4 , it is natural to consider the following question:
Does every ℵ0 -coloring of R have a distinct monochromatic solution to e1 + e2 = e3 + e4 ?
This question can be understood by a bright high school or secondary school student
with no knowledge of vector spaces. The fact that after you show that this question is independent of ZFC you can then abstract the proof to obtain Theorem 8.2 does not diminish
the naturalness of the original question.
9
Acknowledgments
We would like to thank Jacob Fox for references and for writing the paper that pointed us
to this material.
10
References
[1] B. Alexeev and J. Tsimmerman. Equations resolving a conjecture of Rado on partition
regularity. Journal of Combinatorial Theory, Series A, 117:1008–1010, 2010. http:
//www.math.princeton.edu/~jtsimerm.
[2] J. Ceder. Finite subsets and countable decompositions of Euclidean spaces. Revue
Roumaine de Mathematiques Pures et Appliquées, 14(9):1246–1251, 1969.
[3] R. O. Davies. Partioning the plane into denumerably many sets without repeated
differences. Proceedings of the Cambridge Philosophical Society, 72:179–183, 1972.
[4] L. E. Dickson. Lower limit for the number of sets of solutions of xn + y n + z n ≡ 0
(mod p). Journal für die reine und angewandte Mathematik, pages 181–189, 1909.
http://www.cs.umd.edu/~gasarch/res/.
[5] J. Fox. An infinite color analogue of Rado’s theorem. Journal of Combinatorial Theory,
Series A, pages 1456–1469, 2007. http://math.mit.edu/~fox/~publications.html.
[6] W. Gasarch.
Ramsey’s theorem, 2005.
mathnotes/ramsey.pdf.
http://www.cs.umd.edu/~gasarch/
[7] W. Gasarch, C. Kruskal, and A. Parrish. Van der Waerden’s theorem: Variants and
applications. www.gasarch.edu/~gasarch/~vdw/vdw.html.
[8] R. Graham, B. Rothschild, and J. Spencer. Ramsey Theory. Wiley, 1990.
[9] P. Komjáth. Partitions of vector spaces. Periodica Mathematica Hungarica, 28:187–193,
1994.
[10] B. Landman and A. Robertson. Ramsey Theory on the integers. AMS, 2004.
[11] R. Rado. Studien zur Kombinatorik. Mathematische Zeitschrift, pages 424–480, 1933.
http://www.cs.umd.edu/~gasarch/vdw/vdw.html.
[12] R. Rado. Notes on combinatorial analysis. Proceedings of the London Mathematical
Society, pages 122–160, 1943. http://www.cs.umd.edu/~gasarch/vdw/vdw.html.
[13] F. Ramsey. On a problem of formal logic. Proceedings of the London Mathematical
Society, 30:264–286, 1930. Series 2. Also in the book Classic Papers in Combinatorics
edited by Gessel and Rota. The Proceedings of the London Math have put this article
behind a paywall; hence, it will be lost to humanity.
[14] I. Schur. Uber die kongruenz of xm + y m ≡ z m (mod p). Jahresbericht der Deutschen
Mathematiker-Vereinigung, 25:114–116, 1916.
[15] B. van der Waerden. Beweis einer Baudetschen Vermutung. Nieuw Arch. Wisk.,
15:212–216, 1927. This article is in Dutch and I cannot find it online.
11